You are on page 1of 6

Seventh International Conference on CFD in the Minerals and Process Industries

CSIRO, Melbourne, Australia

9-11 December 2009


Department of Mechanical Engineering, Imperial College, London SW7 2AZ, UK
Multiphase Simulation Ltd., London W6 0NB, UK
E-mail address:

This paper reviews recent advances made in the
computational simulation of intermittent flow, and in
particular the slug flow regime in pipes carrying
multiphase fluids. The methodology is based on the
transient, one-dimensional two-fluid model equations that
are solved numerically using fine grids and small time
steps to capture the hydrodynamic instabilities that initiate
waves and slugs. The capture of slugs is achieved in a
completely automatic manner and leads to remarkably
accurate predictions of average slug characteristics when
compared to data from laboratory experiments. More
astonishing is the ability of the one-dimensional model to
capture stochastic nature of slug flow. In industrial
applications involving pipelines extending several
kilometres, special numerical procedures must be adopted
to speed up the computations to keep them within
practical time constraints.

Intermittent flow, especially the slug flow regime, occurs
in many engineering applications, particularly in the
transport of hydrocarbon fluids in pipelines in the oil and
gas industry. The slug flow regime, in which large gas
bubbles flow alternately with liquid slugs at randomly
fluctuating frequency, is usually undesirable since the
intermittency of slugs causes severe adverse conditions.
Firstly, the flow rates of the gas and oil arriving at the
receiving equipment (separators and slug-catchers) can
fluctuate widely thereby undermining the efficacy of the
equipment (e.g. separator flooding); to cater for this
situation, processing facilities are usually designed with
generous safety margins at the expense of cost, weight
and size. Secondly, the flow intermittency results in
highly-unsteady loading on the piping system and
processing equipment, which can result in catastrophic
failure due to metal fatigue. It is therefore important to be
able to predict the onset and subsequent development of
slug flow; it would be even more beneficial if slug
characteristics, such as slug length and frequency could be
calculated as well.



pipe cross-sectional area (m2)

pipe diameter (m)
gas hydraulic diameter (m)
frictional force per unit volume (Nm-3)
friction factor (-)
gravitational acceleration (ms-2)
height of liquid layer (m)
pressure (Nm-2)
Reynolds number (-)
contact perimeter (m)
superficial velocity (ms-1)
velocity (ms-1)
time (s)
distance along the flow (m)
Phase fraction (-)
angle of inclination of pipe to the horizontal (o)
kinematic viscosity (m2s-1)
density (kgm-3)
shear stress (Nm-2)

In horizontal and nearly horizontal pipes, slug flow can be

generated from stratified flow by two main mechanisms:
(i) liquid accumulation due to instantaneous imbalance
between pressure and gravitational forces caused by pipe
undulations, and (ii) natural growth of hydrodynamic
instabilities. In the first, slugs may form at pipe dips due
to the retardation and subsequent accumulation of liquid
in the dips leading to the filling up of the cross-section
with liquid. An extreme example of this terrain induced
slug flow is called severe slugging and occurs when a
slightly inclined pipeline meets a vertical riser (Schmidt
et. al., 1985). This phenomenon is fairly well understood
and methods for calculating its development exist. In the
case of hydrodynamic instabilities, small random
perturbations of short wavelengths arising naturally may
grow into larger and longer waves on the surface of the
liquid. The mechanism behind this growth is the classical
KelvinHelmholtz instability (Lin and Hanratty, 1986).
Such waves may continue to grow picking up liquid
flowing ahead of them, until they bridge the pipe crosssection, thereby forming slugs. These slugs may grow if
the slug fronts travel faster than the tails; conversely they
would collapse. Stable slug flow is obtained if the slug
front and tail travel at the same speed. In real flow, all
these events take place at different times, hence some
slugs grow, others collapse; also slugs may travel at

g gas
i gas-liquid interface
l liquid

Copyright 2009 CSIRO Australia

constitutes a breakthrough in providing a capability (slugcapturing) for predicting the onset of slug flow caused by
hydrodynamic instabilities. It has been demonstrated that
the resulting predictions for the main characteristics of
slug flow compare astonishingly well with experimental
data. Moreover, a remarkable finding in the above cited
works was the generation of slugs of randomly variable
length at different instants in time. Such prediction indeed
tallies with what happens in reality.

different speeds, thereby leading to the merging of some

slugs with others (Taitel and Barnea, 1990).
Slug flow may also arise from both of the aforementioned
mechanisms simultaneously in long pipelines transporting
hydrocarbons. There, slight terrain undulations may lead
to the generation of slugs in addition to those generated by
inherent flow instabilities. In such cases, the slugs
generated from one mechanism interact with those arising
from the other leading to a complex pattern of slugs,
which may overtake and combine.

Despite the fact that the model used is only onedimensional, computations still require an inordinately
long time for real pipelines, since mesh sizes of the order
of few centimetres and time steps of the order of
milliseconds have to be used to capture the instabilities
responsible for slug initiation. It is therefore important to
look at means of accelerating the calculations to render
them of practical use. This paper reviews the situation
with regard to making such applications practical and
presents the results of calculations for a real pipeline.

Oil and gas pipelines are typically several, if not tens of,
kilometres long. The methodology described herein shows
that in order to capture inherent hydrodynamic
instabilities, streamwise mesh sizes of the order of
centimetres and time steps of the order of milliseconds are
needed. To simulate the transient dynamics of the flow in
real pipelines in three dimensions would simply be
infeasible on even the most powerful of current computer
systems; this explains why no work on such simulation
has been attempted in the past apart from short pipe
sections, which is of little use in real pipeline calculations.
One-dimensional, transient models of multiphase flow are
therefore the only practicable means for calculations in
real pipelines. The model utilised for is called the multifluid model in which transport equations for mass,
momentum and energy are formulated for each phase.

The basis of the two-fluid model is the formulation of two
sets of conservation equations for the balance of mass,
momentum and energy for each of the phases. The onedimensional form of the model is obtained by integrating
(area averaging) the flow properties over the crosssectional area of the flow (see Fig. 1).
Transfer of momentum and energy between the walls and
the fluids and between the phases themselves across the
interface is accounted for via source terms in the
equations; they are formulated using empirical
correlations (Ishii and Mishima, 1984).

Dynamic multiphase flow computations in industrial

pipelines are normally based on the two-fluid model, and
some are embodied in well established commercial codes
like OLGA, PROFES and TACITE. All use coarse grids
(pipe segments) of several tens or hundreds of metres in
length; such calculations do not resolve fast transients nor
are able to capture hydrodynamic instabilities. Instead,
two types of approaches are followed to simulate the
effects of intermittent/slug flow, one in which the
presence of slugs is merely represented on an average
basis by empirical closure relations (mainly for shear
forces) pertaining to slug flow. The other approach is
based on the Lagrangian tracking of individual slugs
within the framework of the two-fluid model. Commonly,
slugs of pre-determined length and frequency are assumed
to generate using crude empirical relations. Subsequently,
the position of each slug tail and front is monitored along
the pipe in a Lagrangian manner with time. Neither of the
above approaches can capture the initiation of slug flow
naturally and both rely on empirical slug initiation
relations to start a slug regime calculation. Moreover,
neither is capable of accounting for the effects of
interactions between slug flow and terrain undulations.

The present work is focused on the transport equations for

an isothermal flow. Hence the equations solved are for the
conservation of mass and momentum for the gas and
liquid phases. For one-dimensional stratified and slug flow
they are:
Gas Continuity

( g g ) ( g u g g )
Liquid Continuity

( l l ) ( l ul l )


Gas Momentum

The one-dimensional two-fluid model has been the subject

of extensive analyses of the Kelvin-Helmholtz type to
investigate its ability to simulate wave growth (e.g.
Barnea and Taitel, 1994). Results of these analyses show
clearly that for certain flow conditions, perturbations
occurring in stratified flow are amplified in time resulting
in the formation of waves that may eventually grow to
form slugs. More recently, Issa and Woodburn (1998), and
Issa and Kempf (2003) have shown that by solving
numerically the transient, one-dimensional, two-fluid
model equations it is possible to capture the growth of
disturbances leading to the generation and subsequent
development of slugs in an automatic manner. That work

Copyright 2009 CSIRO Australia


( g g u g ) g g u g2
= g
+ g g g sin (3)
+ Fg + Fi
Liquid Momentum

( l l ul ) l l ul2
= l
( l g ) l g cos
+ l l g sin + Fl Fi

Rels =

In these equations, represents the volume fraction with

the condition that: l+g=1.

DU l


The interfacial friction factor fi is based on the same

correlation of Taitel and Dukler (1976) as for gas-wall
shear. Hence:

f i = C i Re i ni


where Rei is defined as:

Re i =

In the work of Issa and Woodburn (1998), and Issa and

Kempf (2003), the initiation and generation of slug flow
were also shown to be captured as an outcome of the
growth of instabilities in stratified flow in an automatic
manner. The prerequisite for such capability is accurate
numerical resolution in both space and time so that any
small numerical perturbations (due to round-off and other
numerical errors) are captured and allowed to grow
naturally as the governing equations dictate. Figure 2
depicts the outcome of such calculation. In the figure, the
liquid hold-up is plotted along the pipe at several instants
in time starting from stratified flow. The first disturbance
that can barely be discerned in the second inset generates
automatically and then grows to bridge the pipe thereby
generating a slug. Subsequent disturbances generate
periodically to lead to more slugs and eventually resulting
in a continuous train of slugs. What is remarkable is the
fact that the predicted flow is stochastic in nature, in that
no slug is identical to the others, a result that mimics
nature closely.

f g g ug ug S g / A

Fl =


f l L ul ul Sl / A

Fi =


f i g u g ul (u g ul ) Si / A

These friction factors are specified from empirical

correlations. In the present work, the gas-wall factor is
based on the widely used correlation of Taitel & Dukler

f g = C g Re g


where the Reynolds number is based on the gas hydraulic

diameter Dg and defined as:

Re g =



Terms Fg, Fl and Fi in the above equations represent the

frictional forces per unit volume between each phase and
the wall and between the phases at the interface
respectively. They are prescribed by the following closure

In equation (11) the coefficients Ci and ni take the same

values as those for the gas friction factor.

Figure 1: Pipe cross section.

Fg =

Dg (u g ul )

Dg u g


The coefficients Cg and ng take the values of 0.046 and

0.25 respectively if the flow is turbulent (Reg>2100), or 16
and 1 if the flow is laminar (Reg2100).
The best correlation for calculating the liquid-wall friction
factor fl, was found to be that of Spedding & Hand (1997).
Thus, for laminar flow:

fl =



and for turbulent flow:

f l = 0.0262 l Rels



In equations (8) and (9), the Reynolds number Rels is

based on the liquid superficial velocity and the pipe
diameter as:

Copyright 2009 CSIRO Australia

Figure 2: Predicted liquid hold-up distributions along a

pipe at different instants in time in slug flow.

with gas and liquid superficial velocities of 4 and 0.4 m/s,

respectively. The experimental evidence for this case
indicates that the slug lengths are typically in the range of
1230 times the pipe diameter. Also as in real slug flow,
the variations in slug characteristics occur around a
statistical mean, which is what is used to compare the
calculations against the data.

The results of numerous slug capturing computations have

been compared with experimental data in an extensive
validation programme which has proved that the
methodology gives remarkably realistic predictions of
slug flow characteristics (such as slug frequency and
length). An example of that study is presented in figures 3
and 4 (taken from Issa & Kempf, 2003) where the
predicted average slug frequency and average liquid holdup are compared with measurements for horizontal and
slightly downward inclined pipes.

Number of Slugs
Slug Length

Figure 5: Predicted slug length histogram.


Figure 3: Slug frequency in horizontal and slightly

inclined pipes.

The methodology is applied here to a real pipeline 15 km

long and 20 inch diameter carrying oil and gas; the outlet
pressure is 96 bar. The topology of the line is depicted in
Figure 6 where the first section shows the initial 50m,
where the pipe dips after a short horizontal section. The
second section shows the major portion of the pipe
(around 15 km) which is more or less horizontal with few
undulations. The third section depicts the final 50m
section where it rises at an incline and then levels off to
the outlet. A mesh of 400,000 cells was found to give
mesh-independent results.

Figure 4: Mean hold-up in horizontal and inclined pipes.

~50 m

It can be seen from the figures that the agreement between

computations and experiment is excellent bearing in mind
that the predictions are entirely mechanistic without the
need to impose any external perturbation to generate slugs
or any special treatment to simulate their subsequent

~50 m

Figure 6: Topology of pipeline.

Simulations were carried out over nearly two hours of real
flow time starting from arbitrarily assumed initial
conditions; this is the integration time required for the
flow to reach statistically steady state. Time steps of less
than 1 ms were taken and the computations on a single
processor machine require some 45 days to complete. This
clearly is not a practical proposition for industrial
purposes, hence means of accelerating the calculations are
needed and the most effective measure to achieve that
goal is by parallel processing which is being pursued at

A remarkable feature of the computations is that the

predicted slugs exhibit a similar trend to those in real slug
flow in that they are not all of the same length or
frequency. Indeed there is an element of statistical
randomness in their characteristics as is the case in actual
flow. A typical histogram of the slug lengths produced is
shown in Figure 5 (also taken from Issa and Kempf, 2003)
as an example. Such a histogram is obtained from the
computations by monitoring the liquid hold-up at two
points along the pipe (near the outlet) in time. From these
hold-up values the times of arrival and departure of slugs
are established and from these, the slug velocity can be
determined. The length of each slug that passes can thus
be calculated from the passage time and the slug velocity
(in much the same way as is done in actual experiments).
The computed values shown in the figure relate to a case

Copyright 2009 CSIRO Australia

~15 km

Snap-shots of the phase fraction distributions in the three

sections of the pipe are given in panels (a), (b) and (c) in
Figure7. It can be seen that in the first section, the flow is
stratified and is stable. In the middle section, flow
instabilities likely to be induced by either hydrodynamic
instabilities or pipe undulation (or both) develop and form
successive roll waves that continue until the final section

It is clear that in an industrial application where pipelines

are of the order of tens if not hundreds of kilometres in
length, then millions of nodes and time steps may be
necessary to perform the same kind of slug-capturing
calculations shown above. Thus, seemingly simple onedimensional calculations become a major effort requiring
enormous computing effort. In order for the computations
to be completed within practical times to be of use to
industry (hours or days rather than weeks), means of
accelerating the calculations must be found: speed-up of at
least one order of magnitude is necessary to achieve this
In what follows is a general discussion of options that
have been considered for accelerating the calculations. It
will be apparent that only one option (that of parallel
computing) is able to achieve the required speed-up.

of the pipe where after encountering the bend and uphill

section, form slugs that continue until pipe outlet.

Adaptive meshing: The purpose of adaptive meshes is to

minimise the total number of grid nodes in the domain by
concentrating the mesh only in areas where high
resolution is required (i.e. where steep gradients are
present). Although this may appear as a particularly
logical and attractive proposition, it does have its serious
limitations in the context of slug simulation. To begin
with, it is necessary to establish a priori where the relevant
areas in the flow field are in order to concentrate the mesh
in those areas. However, this is extremely difficult if not
impossible in wavy/slug flows where regions of wave or
slug initiation cannot be predicted in advance. The
situation is made even more complicated by the unsteady
nature of the flow where local mesh refinement must be
dynamic in time. Furthermore, experience shows that the
gains from this approach can seldom yield the order of
magnitude improvements sought here.

Figure 7-a: Liquid hold-up distribution in first section

Figure 7-b: Liquid hold-up distribution in middle section

High order differencing schemes: Several high order
schemes for discretising the governing transport
equations, both in space and time, exist. These would
normally need less mesh nodes than lower order schemes
to yield the same degree of numerical accuracy, hence
resulting in less computing time. They are therefore
becoming more widely used. However, here again, the
resulting reduction in number of nodes is nothing like
what is needed to reduce the computational effort
involved in slug capturing. Moreover, in order to capture
the growth of hydrodynamic instabilities, there is an upper
limit on the mesh size to be able to resolve the small
disturbances responsible for initiating the waves
irrespective of the order of the differencing scheme used.
Efficient solution algorithms: Most existing solution
techniques utilise an iterative process to arrive at the final
result because of the strong non-linear coupling of the
multiphase flow equations. In steady state flow, this is
normally the only feasible method of reaching a solution.
In the context of time-dependent calculations, iteration
may be performed at each time step. With so many time
steps needed to march to the final solution, the effort
expended in obtaining the solution therefore may become
enormous. To remedy this, two alternatives may be
adopted. In the first, the multigrid technique can be
utilised to speed up convergence of the iterative process.
However, this comes at the expense of increase in memory
requirement and algorithm complexity as intermediate
solutions on the many grids employed need to be handled.
Also, although reported speed-ups can be achieved on

Figure 7-c: Liquid hold-up distribution in final section

As was mentioned earlier, in order to capture the fine
details of the flow, leading to the growth of hydrodynamic
instabilities that form either roll waves or develop into
slugs, small mesh sizes (grid intervals of few cm) and
short time steps of (of the order of milliseconds) have to
be used. This obviously demands very high computing
resources to perform calculations that commence from
arbitrarily assumed flow conditions (normally stratified
flow), then proceed to capture the initiation and
subsequent development of waves/slugs automatically,
and finally reach a statistically steady state.

Copyright 2009 CSIRO Australia

long pipe was presented. Such computations necessarily

involve hundreds of thousands, and sometimes millions of
nodes as well as millions of time steps in order to capture
and resolve inherent hydrodynamic instabilities. To this
end sufficient computer resources must be made available
for the calculations. It can be argued however that the
magnitude of the problem in wave/slug capturing in a real
pipeline in the oil industry is no bigger than threedimensional cases encountered in other industries. Perhaps
the main additional burden is the time dependent nature of
the slug flow which necessitates the use of transient
computations thereby making them time consuming. A
brief review of the potential benefits and drawbacks of
different techniques to speed up those calculations was
then presented. It is suggested that perhaps the single most
powerful route to reducing computing time is the use of
parallel computing. Here, there are several strategies to
implementing the parallelism, most of which lead to
almost linear speed-up with the number of processors
deployed. Use of such strategies would undoubtedly result
in making slug-capturing simulations manageable as
routine calculations

linear systems of equations, the same may not necessarily

be realised on real non-linear systems. The second
approach is by employing non-iterative techniques
wherein the solution achieved at the end of each time step
contains certain discretisation errors. Such errors
(especially non-conservation of mass) can build up
gradually to the detriment of accuracy of the solution and
stability of the computations, thereby necessitating even
smaller time steps, which defeats the object of the
Parallel computing: With the development of fast
networking technology and multi-core computers,
processors can now be utilised to share the effort in
performing calculations simultaneously. In one approach,
different portions of the problem (e.g. portions of the grid)
are assigned to different processors or computers. In this
way, the execution task is run in parallel on all processors
thereby speeding up the solution process as though it is
run on a much more powerful computer. It is because of
this advantage that parallel computing is relied upon today
in many of the industrial CFD analyses and why the
commercial CFD codes now offer this capability as a main

BARNEA, D. AND TAITEL, Y., (1994), Interfacial
and structural stability of separated flow, Int. J
Multiphase Flow, 20, 387-414.
ISHII, M. AND MISHIMA, K., (1984), Two-fluid
model and hydrodynamic constitutive relations, Nucl.
Eng. Des. 107126.
ISSA R.I. AND KEMPF, M.H.W., (2003), Simulation
of slug flow in horizontal and nearly horizontal pipes with
the two-fluid model, Int. J. Multiphase Flow, 29, 60-95.
ISSA, R. AND WOODBURN, P.J., (1998), Numerical
prediction of instabilities and slug formation in horizontal
two-phase flows, Proc. Third Int. Conf. on Multiphase
Flow, ICMF98, Lyon, France.
LIN, P.Y. AND HANRATTY, T.J., (1986), Prediction
of the initiation of slugs with linear stability theory, Int.
J. Multiphase Flow, 12, 7998.
(1985), Severe slugging in offshore pipeline riser-pipe
systems, J. SPE, 2738.
SPEDDING, P.L. AND HAND, N.P., (1997),
Prediction of stratified gas-liquid co-current flow in
horizontal pipelines, Int. J. Heat Mass Transfer, 40,
TAITEL, Y. AND BARNEA, D., (1990), Two-phase
slug flow, Adv. Heat Transfer, 20, 83132.
TAITEL, Y AND DUKLER, A.E. (1976), A model for
predicting flow regime transitions in horizontal and near
horizontal gasliquid flow. AIChE J., 22, 4755.


Speed Up




20 25


35 40



55 60


70 75


85 90

95 100 105 110

Number of Processors

Figure 8: Example of parallel-computing performance.

Figure 8 shows an example of the speed-up obtained with
parallel computing on several processors and it clearly
illustrates that this is the most feasible way to achieving
the acceleration in the calculations necessary to make the
methodology a practical tool in the oil and gas industry.
The transient one-dimensional two-fluid model has been
shown to be able to simulate wave growth in unstable
stratified flow leading to the initiation and subsequent
evolution of trains of roll waves and slugs. The
methodology has been validated against a range of
laboratory experimental data with excellent agreement
being obtained. The predictions yield stochastic
distribution of slug characteristics such as length and
frequency similar to real slug flow. This ability is rather
surprising, but explicable, in view of the simplicity of the
The method has also been applied to the computation of
intermittent flow in real pipelines; an example of a 15 km

Copyright 2009 CSIRO Australia